This files contains an example of tuning a Logistic Regression model with BayesSearchCV

import pickle
import time

import helpsk as hlp
import pandas as pd
import numpy as np
import plotly.express as px
from sklearn.impute import SimpleImputer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, MinMaxScaler, OneHotEncoder

import plotly.io as pio
pio.renderers.default='notebook'

Load Data

with open('../X_train.pkl', 'rb') as handle:
    X_train = pickle.load(handle)

with open('../y_train.pkl', 'rb') as handle:
    y_train = pickle.load(handle)

hlp.pandas.numeric_summary(X_train)

	# of Non-Nulls	# of Nulls	% Nulls	# of Zeros	% Zeros	Mean	St Dev.	Coef of Var	Skewness	Kurtosis	Min	10%	25%	50%	75%	90%	Max
duration	760	40	5.0%	0	0.0%	21.0	11.7	0.6	1.0	0.6	4.0	9.0	12.0	18.0	24.0	36.0	60.0
credit_amount	800	0	0.0%	38	5.0%	3,203.9	2,932.3	0.9	1.9	3.9	0.0	753.9	1,300.8	2,236.5	3,951.5	7,394.6	18,424.0
installment_commitment	800	0	0.0%	0	0.0%	3.0	1.1	0.4	-0.5	-1.2	1.0	1.0	2.0	3.0	4.0	4.0	4.0
residence_since	800	0	0.0%	0	0.0%	2.9	1.1	0.4	-0.3	-1.4	1.0	1.0	2.0	3.0	4.0	4.0	4.0
age	800	0	0.0%	0	0.0%	35.6	11.4	0.3	1.0	0.7	19.0	23.0	27.0	33.0	42.0	52.0	75.0
existing_credits	800	0	0.0%	0	0.0%	1.4	0.6	0.4	1.3	1.6	1.0	1.0	1.0	1.0	2.0	2.0	4.0
num_dependents	800	0	0.0%	0	0.0%	1.1	0.3	0.3	2.0	2.1	1.0	1.0	1.0	1.0	1.0	2.0	2.0

hlp.pandas.non_numeric_summary(X_train)

	# of Non-Nulls	# of Nulls	% Nulls	Most Freq. Value	# of Unique	% Unique
checking_status	763	37	4.6%	no checking	4	0.5%
credit_history	800	0	0.0%	existing paid	5	0.6%
purpose	800	0	0.0%	radio/tv	10	1.2%
savings_status	800	0	0.0%	<100	5	0.6%
employment	800	0	0.0%	1<=X<4	5	0.6%
personal_status	800	0	0.0%	male single	4	0.5%
other_parties	800	0	0.0%	none	3	0.4%
property_magnitude	800	0	0.0%	car	4	0.5%
other_payment_plans	800	0	0.0%	none	3	0.4%
housing	800	0	0.0%	own	3	0.4%
job	800	0	0.0%	skilled	4	0.5%
own_telephone	800	0	0.0%	none	2	0.2%
foreign_worker	800	0	0.0%	yes	2	0.2%

y_train[0:10]

array([1, 1, 0, 1, 0, 1, 0, 1, 1, 0])

np.unique(y_train, return_counts=True)

(array([0, 1]), array([559, 241]))

np.unique(y_train, return_counts=True)[1] / np.sum(np.unique(y_train, return_counts=True)[1])

array([0.69875, 0.30125])

skopt.BayesSearchCV

https://scikit-optimize.github.io/stable/auto_examples/sklearn-gridsearchcv-replacement.html

search_space = hlp.sklearn_search.LogisticBayesianSearchSpace(random_state=42)

# pip install scikit-optimize
from skopt import BayesSearchCV
#from skopt.space import Real, Categorical, Integer
from sklearn.model_selection import RepeatedKFold

bayes_search = BayesSearchCV(
    estimator=search_space.pipeline(data=X_train),
    search_spaces=search_space.search_spaces(),
    cv=RepeatedKFold(n_splits=5, n_repeats=2, random_state=42),
    scoring='roc_auc',
    n_jobs=-1,
    verbose=1,
    random_state=42,
)

start_time = time.time()
bayes_search.fit(X_train, y_train)
elapsed_time = time.time() - start_time

print(f"Elapsed time to run BayesSearchCV: {elapsed_time:.3f} seconds; {elapsed_time / 60:.1f} minutes")

Elapsed time to run BayesSearchCV: 96.808 seconds; 1.6 minutes

print(bayes_search.best_score_)

0.7698249682368827

print(bayes_search.best_params_)

OrderedDict([('model', LogisticRegression(C=0.057464185559369926, max_iter=1000, random_state=42)), ('model__C', 0.057464185559369926), ('prep__non_numeric__encoder__transformer', OneHotEncoder(handle_unknown='ignore')), ('prep__numeric__imputer__transformer', SimpleImputer()), ('prep__numeric__scaler__transformer', StandardScaler())])

Results

results = hlp.sklearn_eval.MLExperimentResults.from_sklearn_search_cv(
    searcher=bayes_search,                                     
    higher_score_is_better = True,
    parameter_name_mappings = search_space.param_name_mappings()
)

results.to_yaml_file(yaml_file_name = 'Run 1 - Logistic Regression - BayesSearchCV.yaml')

results = hlp.sklearn_eval.MLExperimentResults.from_yaml_file(yaml_file_name = 'Run 1 - Logistic Regression - BayesSearchCV.yaml')

Best Scores/Params

results.best_score

0.7698249682368827

results.best_params

{'model': 'LogisticRegression()',
 'C': 0.057464185559369926,
 'imputer': 'SimpleImputer()',
 'scaler': 'StandardScaler()',
 'encoder': 'OneHotEncoder()'}

results.to_formatted_dataframe(num_rows=100, include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	C	imputer	scaler	encoder
1	0.770	0.751	0.789	0.057	SimpleImputer()	StandardScaler()	OneHotEncoder()
2	0.770	0.750	0.789	0.069	SimpleImputer()	StandardScaler()	OneHotEncoder()
3	0.770	0.750	0.789	0.069	SimpleImputer()	StandardScaler()	OneHotEncoder()
4	0.770	0.750	0.789	0.069	SimpleImputer()	StandardScaler()	OneHotEncoder()
5	0.770	0.750	0.789	0.070	SimpleImputer()	StandardScaler()	OneHotEncoder()
6	0.770	0.750	0.789	0.070	SimpleImputer()	StandardScaler()	OneHotEncoder()
7	0.770	0.750	0.789	0.070	SimpleImputer()	StandardScaler()	OneHotEncoder()
8	0.770	0.750	0.789	0.071	SimpleImputer()	StandardScaler()	OneHotEncoder()
9	0.770	0.750	0.789	0.072	SimpleImputer()	StandardScaler()	OneHotEncoder()
10	0.770	0.750	0.789	0.070	SimpleImputer()	StandardScaler()	OneHotEncoder()
11	0.770	0.750	0.789	0.073	SimpleImputer()	StandardScaler()	OneHotEncoder()
12	0.770	0.750	0.789	0.070	SimpleImputer()	StandardScaler()	OneHotEncoder()
13	0.770	0.750	0.789	0.080	SimpleImputer()	StandardScaler()	OneHotEncoder()
14	0.769	0.750	0.789	0.067	SimpleImputer()	StandardScaler()	OneHotEncoder()
15	0.765	0.735	0.795	0.097	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
16	0.765	0.744	0.785	0.261	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
17	0.765	0.735	0.794	0.106	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
18	0.763	0.733	0.794	0.036	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
19	0.763	0.736	0.790	0.279	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
20	0.760	0.738	0.781	0.009	SimpleImputer()	StandardScaler()	OneHotEncoder()
21	0.758	0.728	0.787	0.005	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
22	0.757	0.738	0.777	<NA>	SimpleImputer()	StandardScaler()	OneHotEncoder()
23	0.756	0.727	0.784	0.001	SimpleImputer(strategy='median')	MinMaxScaler()	OneHotEncoder()
24	0.755	0.727	0.784	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
25	0.755	0.727	0.784	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
26	0.755	0.727	0.784	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
27	0.755	0.727	0.784	0.000	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
28	0.755	0.737	0.773	1.596	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
29	0.754	0.734	0.775	2.107	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
30	0.747	0.730	0.763	11.655	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
31	0.745	0.728	0.762	22.913	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	OneHotEncoder()
32	0.745	0.728	0.762	22.376	SimpleImputer(strategy='median')	MinMaxScaler()	OneHotEncoder()
33	0.744	0.727	0.761	32.731	SimpleImputer()	StandardScaler()	OneHotEncoder()
34	0.742	0.725	0.759	99.507	SimpleImputer()	MinMaxScaler()	OneHotEncoder()
35	0.733	0.703	0.763	0.000	SimpleImputer(strategy='most_frequent')	StandardScaler()	OneHotEncoder()
36	0.731	0.712	0.750	0.003	SimpleImputer(strategy='median')	StandardScaler()	CustomOrdinalEncoder()
37	0.729	0.711	0.748	0.088	SimpleImputer(strategy='most_frequent')	StandardScaler()	CustomOrdinalEncoder()
38	0.728	0.697	0.759	0.000	SimpleImputer()	StandardScaler()	OneHotEncoder()
39	0.728	0.697	0.759	0.000	SimpleImputer(strategy='median')	StandardScaler()	OneHotEncoder()
40	0.725	0.704	0.747	0.000	SimpleImputer()	StandardScaler()	CustomOrdinalEncoder()
41	0.725	0.707	0.743	3.489	SimpleImputer(strategy='median')	StandardScaler()	CustomOrdinalEncoder()
42	0.725	0.707	0.743	6.797	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	CustomOrdinalEncoder()
43	0.725	0.707	0.743	96.252	SimpleImputer(strategy='most_frequent')	StandardScaler()	CustomOrdinalEncoder()
44	0.725	0.707	0.743	99.347	SimpleImputer(strategy='median')	MinMaxScaler()	CustomOrdinalEncoder()
45	0.725	0.706	0.744	0.403	SimpleImputer(strategy='median')	MinMaxScaler()	CustomOrdinalEncoder()
46	0.722	0.700	0.745	0.000	SimpleImputer(strategy='most_frequent')	StandardScaler()	CustomOrdinalEncoder()
47	0.722	0.699	0.744	0.000	SimpleImputer()	StandardScaler()	CustomOrdinalEncoder()
48	0.712	0.689	0.734	0.023	SimpleImputer()	MinMaxScaler()	CustomOrdinalEncoder()
49	0.705	0.682	0.728	0.001	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	CustomOrdinalEncoder()
50	0.699	0.675	0.724	0.000	SimpleImputer()	MinMaxScaler()	CustomOrdinalEncoder()
51	0.699	0.675	0.724	0.000	SimpleImputer()	MinMaxScaler()	CustomOrdinalEncoder()

BayesSearchCV Performance Over Time

results.plot_performance_across_trials().show()

results.plot_performance_across_trials(size=None, color='C').show()

results.plot_performance_across_trials(size='C', color='scaler').show()

---

Variable Performance Over Time

results.plot_parameter_values_across_trials().show()

---

Scatter Matrix

results.plot_scatter_matrix(height=800, width=800 * hlp.plot.GOLDEN_RATIO).show()

---

Variable Performance - Numeric

results.plot_performance_numeric_params()

results.plot_parallel_coordinates().show()

---

Variable Performance - Non-Numeric

results.plot_performance_non_numeric_params()

---

Individual Variable Performance

results.plot_score_vs_parameter(
    parameter='C',
    color='scaler'
)

results.plot_score_vs_parameter(
    parameter='C',
    color='encoder'
)

---

Regression on roc_auc Mean

score_variable = results.primary_score_name + ' Mean'

score_dataframe = results.to_dataframe()
score_dataframe = score_dataframe.drop(columns=[x for x in score_dataframe.columns
                              if x not in [score_variable] + results.parameter_names])
score_dataframe.head()

	roc_auc Mean	C	imputer	scaler	encoder
19	0.769825	0.057464	SimpleImputer()	StandardScaler()	OneHotEncoder()
21	0.769730	0.068645	SimpleImputer()	StandardScaler()	OneHotEncoder()
34	0.769730	0.068655	SimpleImputer()	StandardScaler()	OneHotEncoder()
33	0.769694	0.069019	SimpleImputer()	StandardScaler()	OneHotEncoder()
25	0.769680	0.069805	SimpleImputer()	StandardScaler()	OneHotEncoder()

cleaned_column_names = [''.join(e for e in x.replace(' ', '_') if e == '_' or e.isalnum()) for x in score_dataframe.columns.tolist()]
cleaned_column_names = dict(zip(score_dataframe.columns.tolist(), cleaned_column_names))
cleaned_column_names

{'roc_auc Mean': 'roc_auc_Mean',
 'C': 'C',
 'imputer': 'imputer',
 'scaler': 'scaler',
 'encoder': 'encoder'}

score_dataframe = score_dataframe.rename(columns=cleaned_column_names)

import statsmodels.formula.api as smf

y_column = 'roc_auc_Mean'
X_columns = score_dataframe.columns.tolist()
X_columns.remove(y_column)
X_columns = hlp.string.collapse(X_columns, separate=" + ", surround="")

formula = f"{y_column} ~ {X_columns}"
print(formula)
model = smf.ols(formula=formula, data = score_dataframe)
results = model.fit()
print(results.summary())

roc_auc_Mean ~ C + imputer + scaler + encoder
                            OLS Regression Results                            
==============================================================================
Dep. Variable:           roc_auc_Mean   R-squared:                       0.728
Model:                            OLS   Adj. R-squared:                  0.698
Method:                 Least Squares   F-statistic:                     23.61
Date:                Mon, 14 Feb 2022   Prob (F-statistic):           1.88e-11
Time:                        18:32:51   Log-Likelihood:                 154.15
No. Observations:                  50   AIC:                            -296.3
Df Residuals:                      44   BIC:                            -284.8
Df Model:                           5                                         
Covariance Type:            nonrobust                                         
======================================================================================================================
                                                         coef    std err          t      P>|t|      [0.025      0.975]
----------------------------------------------------------------------------------------------------------------------
Intercept                                              0.7188      0.005    153.899      0.000       0.709       0.728
imputer[T.SimpleImputer(strategy='median')]           -0.0034      0.005     -0.641      0.525      -0.014       0.007
imputer[T.SimpleImputer(strategy='most_frequent')]    -0.0033      0.004     -0.843      0.404      -0.011       0.005
scaler[T.StandardScaler()]                             0.0063      0.004      1.767      0.084      -0.001       0.013
encoder[T.OneHotEncoder()]                             0.0375      0.004      9.553      0.000       0.030       0.045
C                                                  -3.878e-05   7.28e-05     -0.533      0.597      -0.000       0.000
==============================================================================
Omnibus:                       18.462   Durbin-Watson:                   0.841
Prob(Omnibus):                  0.000   Jarque-Bera (JB):               22.437
Skew:                          -1.466   Prob(JB):                     1.34e-05
Kurtosis:                       4.476   Cond. No.                         104.
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

from sklearn.preprocessing import StandardScaler
from sklearn.compose import ColumnTransformer
import pandas as pd

scaler = StandardScaler()
#scaler.fit_transform(bayes_search_df)

numeric_columns = hlp.pandas.get_numeric_columns(score_dataframe)
non_numeric_columns = hlp.pandas.get_non_numeric_columns(score_dataframe)

print(numeric_columns)
print(non_numeric_columns)

numeric_pipeline = Pipeline([
    ('scaling', StandardScaler()),
])

transformations_pipeline = ColumnTransformer([
    ('numeric_pipeline', numeric_pipeline, numeric_columns),
    ('non_numeric_pipeline', 'passthrough', non_numeric_columns)
])

score_dataframe_transformed = transformations_pipeline.fit_transform(score_dataframe)
score_dataframe_transformed = pd.DataFrame(score_dataframe_transformed,
                                           columns= numeric_columns + non_numeric_columns)
score_dataframe_transformed.head()

['roc_auc_Mean', 'C']
['imputer', 'scaler', 'encoder']

	roc_auc_Mean	C	imputer	scaler	encoder
0	1.04166	-0.335822	SimpleImputer()	StandardScaler()	OneHotEncoder()
1	1.037159	-0.335351	SimpleImputer()	StandardScaler()	OneHotEncoder()
2	1.037159	-0.33535	SimpleImputer()	StandardScaler()	OneHotEncoder()
3	1.035465	-0.335335	SimpleImputer()	StandardScaler()	OneHotEncoder()
4	1.034805	-0.335302	SimpleImputer()	StandardScaler()	OneHotEncoder()

score_dataframe_transformed['roc_auc_Mean'] = score_dataframe_transformed['roc_auc_Mean'].astype('float')
score_dataframe_transformed['C'] = score_dataframe_transformed['C'].astype('float')

print(formula)
model = smf.ols(formula=formula,
                data = score_dataframe_transformed)
results = model.fit()
print(results.summary())

roc_auc_Mean ~ C + imputer + scaler + encoder
                            OLS Regression Results                            
==============================================================================
Dep. Variable:           roc_auc_Mean   R-squared:                       0.728
Model:                            OLS   Adj. R-squared:                  0.698
Method:                 Least Squares   F-statistic:                     23.61
Date:                Mon, 14 Feb 2022   Prob (F-statistic):           1.88e-11
Time:                        18:32:51   Log-Likelihood:                -38.745
No. Observations:                  50   AIC:                             89.49
Df Residuals:                      44   BIC:                             101.0
Df Model:                           5                                         
Covariance Type:            nonrobust                                         
======================================================================================================================
                                                         coef    std err          t      P>|t|      [0.025      0.975]
----------------------------------------------------------------------------------------------------------------------
Intercept                                             -1.3880      0.217     -6.403      0.000      -1.825      -0.951
imputer[T.SimpleImputer(strategy='median')]           -0.1625      0.253     -0.641      0.525      -0.673       0.348
imputer[T.SimpleImputer(strategy='most_frequent')]    -0.1546      0.183     -0.843      0.404      -0.524       0.215
scaler[T.StandardScaler()]                             0.2966      0.168      1.767      0.084      -0.042       0.635
encoder[T.OneHotEncoder()]                             1.7766      0.186      9.553      0.000       1.402       2.151
C                                                     -0.0436      0.082     -0.533      0.597      -0.208       0.121
==============================================================================
Omnibus:                       18.462   Durbin-Watson:                   0.841
Prob(Omnibus):                  0.000   Jarque-Bera (JB):               22.437
Skew:                          -1.466   Prob(JB):                     1.34e-05
Kurtosis:                       4.476   Cond. No.                         5.98
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.

coefficients = pd.DataFrame({
    'feature': results.params.index,
    'coefficient': results.params,
    'p_value': results.pvalues,
})
coefficients = coefficients.query("feature != 'Intercept'")
coefficients['Stat Sig'] = coefficients['p_value'] <= 0.05
coefficients

	feature	coefficient	p_value	Stat Sig
imputer[T.SimpleImputer(strategy='median')]	imputer[T.SimpleImputer(strategy='median')]	-0.162479	5.247883e-01	False
imputer[T.SimpleImputer(strategy='most_frequent')]	imputer[T.SimpleImputer(strategy='most_frequen...	-0.154624	4.039380e-01	False
scaler[T.StandardScaler()]	scaler[T.StandardScaler()]	0.296624	8.408081e-02	False
encoder[T.OneHotEncoder()]	encoder[T.OneHotEncoder()]	1.776633	2.686761e-12	True
C	C	-0.043564	5.967535e-01	False

score_variable

'roc_auc Mean'

px.bar(
    data_frame=coefficients.reindex(coefficients['coefficient'].abs().sort_values(ascending=True).index),
    y='feature',
    x='coefficient',
    color='Stat Sig',
    title=f"Regression Coefficients of Hyper-parameters against '{score_variable}'",
    height=600,
    width=600*hlp.plot.GOLDEN_RATIO
)

Feature Importance

https://scikit-learn.org/stable/auto_examples/ensemble/plot_forest_importances.html

from sklearn.inspection import permutation_importance

forest = bayes_search.best_estimator_['model']

start_time = time.time()
result = permutation_importance(
    bayes_search.best_estimator_, X_train, y_train, n_repeats=10, random_state=42, n_jobs=2
)
elapsed_time = time.time() - start_time
print(f"Elapsed time to compute the importances: {elapsed_time:.3f} seconds")

feature_names = X_train.columns.to_list()
forest_importances = pd.Series(result.importances_mean, index=feature_names)
forest_importances = forest_importances.sort_values(ascending=False)

Elapsed time to compute the importances: 4.851 seconds

import matplotlib.pyplot as plt
fig, ax = plt.subplots()
forest_importances.plot.bar(yerr=result.importances_std, ax=ax)
ax.set_title("Feature importances using permutation on full model")
ax.set_ylabel("Mean accuracy decrease")
fig.set_size_inches(9, 6)
fig.tight_layout()
plt.show()

temp = X_train.copy()
temp['default'] = y_train
temp.groupby('foreign_worker').agg({'default': np.mean})

	default
foreign_worker
yes	0.308290
no	0.107143

fig = px.box(
    data_frame=temp,
    y='age',
    x='default',
#    size=size_variable,
#    color=color_variable,
#    trendline='lowess',
#     labels={
#         score_variable: f"Average Cross Validation Score ({parser.primary_score_name})",
#     },
#     title=f"<b>{x_variable}</b> - Performance<br>" \
#           f"<sup>Size of point corresponds to '{size_variable}'</sup>",
#     custom_data=['labels'],
    height=600,
    width=600*hlp.plot.GOLDEN_RATIO
)
fig.show()

NOTE: foreign worker seems like it should be important but is ranked last in feature importance.